Peter Anderson

There are some comments about how to get around the memory requirements in issue #3

I'm afraid none of this is tested on windows, you might have to do your own research on this one, sorry!

See https://github.com/peteanderson80/bottom-up-attention#pretrained-features

There is an alternative download option now using Google cloud storage. It is user pays but it is very fast and should work much better, even with everything in a...

The mean and weighted mean numbers should be much closer than your results - the only difference is correction for class imbalance. Are you still having an issue with this?

Hi @jwyang, Sorry I haven't responded sooner. We did not exclude zeros in our calculation. It seems like there is some difference in the validation set that is being used,...

Hi @yuzcccc, I don't have the original log file, but I've added an [example log file](https://github.com/peteanderson80/bottom-up-attention/blob/09f8384031dc5ec145a18973a22cb58b5dc0e311/experiments/logs/faster_rcnn_end2end_final_ResNet-101_.txt.2017-06-29_16-27-08) from training with a single gpu for 16K iterations, which should give some indication...

Thanks @jwyang for investigating. I have shared our pickled datasets so you can see if you get the same: - [vg_1600-400-20_train_gt_roidb.pkl](https://storage.googleapis.com/bottom-up-attention/vg_1600-400-20_train_gt_roidb.pkl) - [vg_1600-400-20_val_gt_roidb.pkl](https://storage.googleapis.com/bottom-up-attention/vg_1600-400-20_val_gt_roidb.pkl)

There is some code in our caption model repo for this: https://github.com/peteanderson80/Up-Down-Captioner/blob/master/scripts/demo.ipynb

No, I don't think the object and attribute labels are in the tsv files. In hindsight, we should have included them. The demo above is end-to-end, i.e. running Faster R-CNN...